ReviewMind

Problem Code review is one of the most time-consuming and error-prone parts of the backend development workflow. On FastAPI teams, reviewers frequently miss three critical categories of issues: 1. Breaking API changes that silently break clients 2. Common backend pitfalls (missing response models, blocking calls in async endpoints, improper dependency injection, session leaks, etc.) 3. Regressions — bugs that were fixed in the past but get reintroduced Because institutional knowledge lives only in old pull requests, closed issues, and scattered review comments, the same mistakes keep happening. Manual review is slow, inconsistent, and heavily dependent on who is available that day. Solution We built a history-aware Pull Request reviewer specifically for FastAPI teams. When a PR is opened, the system: • Receives the GitHub webhook and parses the diff • Runs three specialized lenses in parallel: – API Contract Lens: detects breaking schema and endpoint changes – Backend Pitfall Lens: checks for common FastAPI anti-patterns and project rule violations – Regression Lens: uses a RAG store of past PRs, bug fixes, and review comments to find similar historical issues • Sends the findings plus retrieved historical context to IBM Bob • Bob explains each problem in plain language and suggests concrete fixes, citing similar past cases • Posts a single, well-structured comment on the PR with an overall risk score Target Users Backend engineers and tech leads working on FastAPI services who want faster, more consistent, and more knowledgeable code reviews. How Users Interact Developers simply open a Pull Request as usual. Within seconds they receive an automated review comment that highlights real risks, explains why they matter, and shows how the team solved similar problems before. No extra tools or browser extensions are required. What Makes It Unique Most AI code review tools only look at the current diff. Our solution adds organizational memory. By combining static analysis, project-specific rules, and retrieval over the team’s own history, it catches regressions and repeated mistakes that pure LLM reviewers miss. Using IBM Bob’s Agent mode, subagents, and document understanding allows the system to orchestrate the full multi-step workflow rather than just generating code suggestions. Impact The system reduces manual review effort, lowers the chance of shipping regressions and breaking changes, and shortens the feedback loop from hours to seconds — while teaching the team from its own past.